Substrate Engineering is the engineering discipline concerned with producing deterministic interpretation from inherently ambiguous inputs. Rather than constraining model behavior after the fact, it establishes explicit meaning before execution.
Once meaning is governed and consistently interpreted, independently trained models can produce SHA-256 byte-identical reproducible outputs for the same input.
The goal is not consistency within a single model. The goal is reproducibility across models, runs, environments, organizations, and time.
Reliable decisions emerge from stable interpretation.
Modern software relies on protocols to govern data transport, identity, and execution. AI systems, however, still lack equivalent infrastructure for governing interpretation itself.
Engineering already treats data, identity, and state as first-class primitives, each has standards that make it dependable. Meaning never received the same treatment.
In most agentic systems, how an instruction is interpreted is left as an implementation detail, so the same input can drift across several equally valid readings. Substrate Engineering addresses this gap by treating meaning as a first-class engineering primitive rather than an emergent property of model behavior.
Instead of relying on retries, guardrails, or increasingly complex orchestration layers to compensate for unreliable outcomes, Substrate Engineering focuses on the conditions that produce reliable outcomes in the first place.
The key insight is that determinism and creativity are not in conflict. They operate in different domains. A language model is free to interpret, phrase, and reason in whatever way its training enables. This is exactly where probabilistic systems excel. Substrate Engineering pins the decision on that interpretation, not the interpretation itself.
The result is deterministic decision-making on top of probabilistic systems.
TCP/AP is an implementation of these principles: an open protocol for Deterministic AI, resolving meaning reliably in non-deterministic systems so AI decisions stay reproducible across models, runs, and environments.
Where other approaches add layers of complexity around model output (retries, consensus, evaluators), TCP/AP changes what a model is asked to produce. Instead of asking a model to make a decision, you ask it to produce evidence. The protocol makes the decision.